Abstract:

Through the research and analysis of the original SSD（Single Shot Multibox Detector） model, the problem of its weak ability to detect small targets is solved. An improved model based on dense module and feature fusion is proposed. Based on Inception-ResNet-V2 and DenseNet, this model absorbs the research ideas of sparse connection in inception module and dense connection in dense network, and fuses the two methods together to propose the feature extraction structure of Inception-Dense. In the part of multi-scale detection, the feature fusion module of feature pyramid is used for reference and improved to strengthen the detection ability of small and medium-sized targets. The mapping mechanism of the default box is also reset according to the characteristics of the improved model and experimental data set. The results show that the average test accuracy（mAP） of this method on the Kitti dataset is 83.8%; the recognition rate is about 11 percentage points higher than the 72.8% of the original SSD model. FPS also has a nearly 38% improvement, from 39 to 54.

Key words: deep learning, SSD（Single Shot Multibox Detector）, target detection, neural network

摘要：

通过对原SSD（Single Shot Multibox Detector）模型的研究与分析，针对其对小目标检测能力较弱的问题，提出了一种基于密集模块与特征融合操作的改进模型。该模型以Inception-ResNet-V2与DenseNet为基础，吸取了inception模块中稀疏连接与密集网络中密集连接的研究思路，将两种方法融合在一起，提出了Inception-Dense特征提取结构。在多尺度检测的部分，借鉴并改进了特征金字塔的特征融合模块来加强对中小目标的检测能力。根据改进模型及实验数据集的相关特性，对默认框的映射机制也进行了重新设定。结果表明：该方法在Kitti数据集上的平均测试精确度（mAP）为83.8%；识别率相比于原SSD模型的72.8%，提升了11个百分点。FPS方面也有接近38%的提升，从原来的39提升到了54。

关键词: 深度学习, SSD（Single Shot Multibox Detector）, 目标检测, 神经网络